How well do we melt snow? A multi-model intercomparison project (MuMIP) illustrates the challenges of hydrological prediction

Abstract:

The loss of seasonal snow from climate change will have substantial impacts on hydrological and ecological systems. Previous snow model inter-comparisons focused on consistency with observations but have not assessed the sensitivity of snow models to future climate perturbations. We train a multi-model that mimics common land surface model assumptions on historical observations and show that a range of reasonable model decisions and parameter combinations can accurately predict snowpack’s mass and energy balance at a well-instrumented site in southwestern Colorado, USA. These models are then perturbed with two pseudo global warming scenarios with and without changes to precipitation (PGW and PGW₀, respectively) to produce an ensemble of snowpack projections from a spectrum of plausible model assumptions. Ensembles from both future scenarios agree on projected decreases in maximum SWE and snow retention, while snow melt rate increases or decreases depending on model assumptions such as albedo approximations. We constrain the projected ensemble spread by including snowpack melt rate and cold content as additional objective functions, which reduces the number of historically-accurate simulations, used for future predictions. Constraining the ensemble spread on historical cold content and melt rate results in a decreased maximum SWE by 45% and 35%, and an increased melt rate by 9% and 12% under PGW and PGW₀ scenarios, respectively. However, the ensemble spread of only max SWE objective function shows 40% and 29% decreases in Max SWE, and 8% and 11% decreases in melt rate under PGW and PGW₀ scenarios, respectively. Our results suggest that equifinality in snow model predictions may lead to poor assumptions for earth system models and potentially misrepresenting future hydrology. We demonstrate that there is a 50% chance that a historically-accurate snow model has more than 6% errors in future SWE prediction. Incorporating more observational consolation, like snowmelt rate and snowpack cold content, can constrain future predictions of SWE by 55% and melt rate by 43%. These results call into questioning the underlying parsimony of physically-based models and provide a multi-model framework to explore the sensitivity of hydrologic simulations to climate change projections in these complex models that often lack parsimony.